A transferable Boltzmann Generator for small molcules conformations
||A transferable Boltzmann Generator for small molcules conformations|
||Juan Viguera Diez <firstname.lastname@example.org>|
||Chalmers tekniska högskola|
||2022-10-01 – 2023-04-01|
The problem drug-discovery can be compared to finding a needle in a haystack. Among the 10^60-10^100 theoretically possible drug-like compounds, the aim is to find molecules with a set of properties (such as solubility or low toxicity). Machine learning has shown promising results for the task of efficiently exploring the vast chemical space searching for candidate compounds. However, most of the proposed methods do not consider the 3D structure of molecules, which strongly influences some of their properties, and sampling molecular conformations remains to be a challenging problem. In this project, we aim to design a machine learning model that is able to efficiently generate physically realistic 3D structures of drug-like molecules, overcoming limitations from traditional methods such as poor mixing and low acceptance ratio. To do that we aim to use deep generative models in a similar way as they have previously been used to generate realistic images or music. Specifically, we plan using Boltzmann Generators, which are generative models able to learn from both examples of 3D molecular conformations and from an energy model.
For more info please visit our publication at https://moleculediscovery.github.io/workshop2021/, https://cloud.ml.jku.at/s/sKtfdFpoTp9F7sJ